Return to search

Modeling Temporal Bias of Uplift Events in Recommender Systems

Today, commercial industry spends huge amount of resources in advertisement campaigns, new marketing strategies, and promotional deals to introduce their product to public and attract a large number of customers. These massive investments by a company are worthwhile because marketing tactics greatly influence the consumer behavior. Alternatively, these advertising campaigns have a discernible impact on recommendation systems which tend to promote popular items by ranking them at the top, resulting in biased and unfair decision making and loss of customers’ trust. The biasing impact of popularity of items on recommendations, however, is not fixed, and varies with time. Therefore, it is important to build a bias-aware recommendation system that can rank or predict items based on their true merit at given time frame.
This thesis proposes a framework that can model the temporal bias of individual items defined by their characteristic contents, and provides a simple process for bias correction. Bias correction is done either by cleaning the bias from historical training data that is used for building predictive model, or by ignoring the estimated bias from the predictions of a standard predictor. Evaluated on two real world datasets, NetFlix and MovieLens, our framework is shown to be able to estimate and remove
the bias as a result of adopted marketing techniques from the predicted popularity of
items at a given time.

Identiferoai:union.ndltd.org:kaust.edu.sa/oai:repository.kaust.edu.sa:10754/292299
Date08 May 2013
CreatorsAltaf, Basmah
ContributorsZhang, Xiangliang, Computer, Electrical and Mathematical Sciences and Engineering (CEMSE) Division, Gao, Xin, Moshkov, Mikhail
Source SetsKing Abdullah University of Science and Technology
LanguageEnglish
Detected LanguageEnglish
TypeThesis
Rights2016-03-16, At the time of archiving, the student author of this thesis opted to temporarily restrict access to it. The full text of this thesis became available to the public after the expiration of the embargo on 2016-03-16.

Page generated in 0.0021 seconds